Scrap and rework occupy an odd position in manufacturing accounts. They’re measured, reported, and largely tolerated — treated as a rate to be managed rather than a set of specific causes to be eliminated. Part of the reason is that by the time a defect is detected, the information needed to explain it has usually dispersed.

Inspection finds defects; it doesn’t prevent them

End-of-line inspection is detection after the fact. The material is consumed, the machine time is spent, and the choice is between scrapping and reworking. Both are pure loss.

The cost is also understated in most reporting. A scrapped unit costs material plus all labour and machine time invested, plus the capacity consumed on the constraint, plus the replacement production, plus any delivery delay that follows. Reported scrap cost frequently captures only the material.

Predicting rather than detecting

Defects usually have precursors in data you already collect. Process parameter drift — temperature, pressure, speed, cycle time moving within tolerance but away from where good output has historically occurred. Material batch effects, where certain supplier lots correlate with higher defect rates. Machine condition, where a tool approaching end of life produces subtle dimensional drift before it produces rejects. Environmental factors like humidity for sensitive processes. Combinations, where two conditions individually acceptable produce failure together.

That last category is where analysis earns its keep, because interaction effects are close to impossible to identify by eye and routinely explain the defects nobody can account for.

Acting early

With a prediction, several responses become available before the defect exists. Adjust process parameters back toward the historical good zone. Change tooling at the point where drift begins rather than when rejects appear. Route a suspect material batch to a less sensitive product. Increase inspection sampling temporarily rather than permanently.

Each of these costs a fraction of the scrap it prevents, and each requires only a few hours of warning.

Data prerequisites

This is the honest constraint. Predictive quality needs process parameters recorded against production runs, defects recorded with enough specificity to distinguish types rather than a single reject count, material batch linkage through to output, and machine and tooling identification per run.

Many manufacturers have some of this and not all. The commonest gaps are defect type granularity — everything recorded as reject with no mode — and batch traceability. Both are fixable, and fixing them is the actual first project regardless of what analytics follows.

Where to start

Take your highest-cost defect mode on your highest-volume product, and gather what data exists around it for six months. Look for correlation with process parameters, material batches, shifts, and equipment before building anything more sophisticated. A meaningful share of quality problems resolve at this stage, with no modelling at all — the analysis simply reveals a pattern nobody had assembled before.

Involve the operators in reading the results. They frequently recognise a pattern immediately that would take an analyst weeks to interpret, because they know what changed on the line in the period the data covers.

ticktick.ai links material batches, process data, and defect records through production, so correlations surface without manual data assembly.